Real-Time Deadlines Reveal Temporal Awareness Failures in LLM Strategic Dialogues

TL;DR

The paper investigates whether LLM-based agents can internally represent and act on continuous time during real-time, multi-turn negotiations, revealing a systematic temporal-awareness deficit. It introduces a controlled bilateral negotiation framework with Time-Limit-Only and Time-Aware conditions, along with a latency model, to disentangle temporal tracking from strategic competence across multiple models and two scenarios. Key findings show that explicit remaining-time feedback substantially boosts deal closure and acceptance, while internal time tracking fails under real-time deadlines; turning to discrete turn limits yields near-perfect performance, underscoring the specific challenge is continuous-time reasoning rather than general bargaining ability. The results generalize across scenarios and model families, underscoring a practical limit for deploying time-sensitive LLM agents and motivating architectural or training changes to embed temporal representations and time-aware policies.

Abstract

Large Language Models (LLMs) generate text token-by-token in discrete time, yet real-world communication, from therapy sessions to business negotiations, critically depends on continuous time constraints. Current LLM architectures and evaluation protocols rarely test for temporal awareness under real-time deadlines. We use simulated negotiations between paired agents under strict deadlines to investigate how LLMs adjust their behavior in time-sensitive settings. In a control condition, agents know only the global time limit. In a time-aware condition, they receive remaining-time updates at each turn. Deal closure rates are substantially higher (32\% vs. 4\% for GPT-5.1) and offer acceptances are sixfold higher in the time-aware condition than in the control, suggesting LLMs struggle to internally track elapsed time. However, the same LLMs achieve near-perfect deal closure rates ($\geq$95\%) under turn-based limits, revealing the failure is in temporal tracking rather than strategic reasoning. These effects replicate across negotiation scenarios and models, illustrating a systematic lack of LLM time awareness that will constrain LLM deployment in many time-sensitive applications.

Figures (5)

• Figure 1: Experimental Design and Study Flow. Two LLM agents negotiate over a multi-issue hiring package under varying levels of time pressure (240, 300, or 360 seconds), assuming a fixed generation latency of 150 WPM (words per minute). In the Control Condition, agents are informed only of the total time available. In the Time-Aware Condition, agents receive both the total time limit and explicit updates on the remaining time at each conversational turn. Each negotiation ends when agents reach agreement, take their outside option (BATNA), or exhaust their allotted time or turns.
• Figure 2: Deal Closure by Treatment and Time Limit. Deal closure rates for GPT-5.1-chat-latest agents across experimental conditions and time limits. Each bar represents the proportion of negotiations reaching successful deals (N=100 per condition). The Time-Aware condition, where agents receive explicit temporal feedback at each turn, shows higher closure rates than the Control condition across all time limits. Error bars represent 95% Wilson confidence intervals. Results suggest that LLMs cannot reliably internalize time pressure without explicit external signals.
• Figure 3: Urgency ablation: Deal closure by treatment and time limit. The Urgency advantage over Time-Aware is significant at $T{=}360$s ($p<0.001$) and significant when aggregating across time limits. Urgency's advantage demonstrates models can respond to deadline pressure when explicitly cued, but fail to generate such urgency signals internally (Control). Error bars show 95% Wilson confidence intervals (N=100 negotiations per condition per time limit).
• Figure A.1: Rubbermind - Deal Closure by Treatment and Time Limit. Deal closure rates for GPT-5.1-chat-latest agents across experimental conditions and time constraints in a second, more difficult negotiation scenario (Rubbermind). Each bar represents the proportion of negotiations reaching successful agreement (N=100 per condition). The Time Aware condition, where agents receive explicit temporal feedback at each turn, shows higher closure rates than the Control condition across 360-480s time limits. Error bars represent 95% Wilson confidence intervals. Results demonstrate that LLMs struggle to reliably internalize time pressure without explicit external signals.
• Figure A.2: Deal Closure by Treatment and Time Limit without Latency Factor. Deal closure rates for GPT-5.1-chat-latest agents across experimental conditions and time constraints with the 150 word per minute latency adjustment removed. Each bar represents the proportion of negotiations reaching successful agreement (N=100 per condition). The Time Aware condition, where agents receive explicit temporal feedback at each turn, shows higher closure rates for short and medium time limits. For higher time limits, both conditions approach ceiling performance. Error bars represent 95% Wilson confidence intervals. Results demonstrate the temporal-awareness effect is not an artifact of the words-per-minute timing assumption.